Single-cell dissection of persistent tumor antigens in non-responders reveals opportunities for TAA-targeted vaccination after neoadjuvant therapy in esophageal squamous cell carcinoma

Abstract

Background

Neoadjuvant immune checkpoint blockade (ICB) combined with chemotherapy has demonstrated higher response rates than chemotherapy alone in resectable esophageal squamous cell carcinoma (ESCC). Although tumor microenvironment (TME) features associated with treatment response have been studied in the contexts of both chemotherapy and immunotherapy, direct comparisons of these characteristics and their implications for novel therapeutic development remain largely unexplored.

Methods

Paired pretreatment and post-treatment tumor samples were prospectively collected from 55 patients with locally advanced ESCC enrolled in a multicenter, phase 3 clinical trial (ChiCTR2000040034). Patients were randomized to receive neoadjuvant chemotherapy (paclitaxel with cisplatin, TP; n=14) or its combination with PD-1 blockade (camrelizumab, albumin-bound paclitaxel and cisplatin/paclitaxel and cisplatin, Cam+nab-TP/TP; n=41), followed by surgical resection. Pathological response was defined based on Mandard’s Tumor Regression Grade scoring system. Changes in immune cell populations and tumor-associated antigens (TAAs) within the TME in relation to response were characterized using single-cell RNA and T cell receptor sequencing.

Results

Both chemotherapy and chemoimmunotherapy responders exhibited shared temporal dynamics including dendritic cell remodeling, cytotoxic CD8⁺ T cell contraction, and memory T cell expansion, with chemoimmunotherapy uniquely suppressing the clonal expansion of GZMB⁺TIGIT⁺ T cells. In contrast, non-responders failed to elicit effective antitumor immunity and displayed persistent expression of targetable TAAs, fully preserved HLA machinery, and clonal expansion of dysfunctional GZMB⁺TIGIT⁺ T cells.

Conclusions

Our study identifies key immune contributors correlated with response to neoadjuvant therapies and a panel of TAAs in non-responders. These findings support the development of TAA-targeted therapeutic vaccines and combination strategies incorporating ICB to overcome resistance in non-responders.

WHAT IS ALREADY KNOWN ON THIS TOPIC

• A deeper understanding of how the tumor microenvironment (TME) dictates therapeutic response is critical for developing tailored antitumor strategies. However, identifying immune cell subsets that distinctly mediate the efficacy of chemotherapy or immunotherapy is particularly difficult in the neoadjuvant setting, where real-world patients often receive combination regimens, limiting advances toward more effective esophageal squamous cell carcinoma (ESCC) therapies.

WHAT THIS STUDY ADDS

• This study leverages a homogeneous cohort with rigorously standardized treatment regimens to identify immune cell subsets associated with response to neoadjuvant chemotherapy with or without immune checkpoint blockade (ICB) and highlights the critical importance of active tumor antigen-specific immune responses in both regimens. Importantly, we provide evidence that tumor-associated antigens (TAAs) persist in non-responders over the course of therapy. These findings not only support the rationale for developing TAA-targeted immunotherapies for patients receiving either regimen, in particular chemotherapy, but also demonstrate the necessity of incorporating ICB to reinvigorate exhausted T cells within combination strategies.

HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY

• To our knowledge, this is the first study in ESCC to delineate agent-specific TME alterations within neoadjuvant combination regimens and to demonstrate persistent TAA expression in non-responders at the single-cell resolution, providing key insights for rational drug development and the optimization of future combination strategies.

Introduction

Standard of care treatment options for resectable locally advanced esophageal squamous cell carcinoma (LA-ESCC) include chemoradiotherapy or chemotherapy, followed by esophagectomy, yet long-term survival rates remain low.^{1 2} Immune checkpoint blockade (ICB) immunotherapy has revolutionized therapy for advanced or metastatic ESCC by extending patient survival, prompting its incorporation into the neoadjuvant setting.3,8 Our recent randomized phase 3 clinical trial⁹ showed that two regimens of neoadjuvant anti-PD-1 plus chemotherapy achieved markedly higher pathological complete response (pCR) rates than chemotherapy alone (28.0% and 15.4% vs 4.7%), consistent with multiple phase 1b and 2 trials reporting pCR rates of 24.1% to 39.2%.10,13 However, two key questions remain unresolved: (1) a substantial proportion of patients fail to achieve pCR and (2) the validity of pCR as a surrogate endpoint remains under debate even in the context of ICB, as no data to date demonstrate that neoadjuvant ICB improves event-free survival, disease-free survival, or overall survival.^{14 15} These unresolved questions underscore the need to elucidate the immune contributors that drive active and durable antitumor responses to enable the development of more effective therapeutics.

Multiple lines of evidence indicate that treatment can elicit immune responses against tumor antigens, helping explain why immunological features of the tumor microenvironment (TME), particularly CD8⁺ tumor-infiltrating lymphocytes, are associated with favorable prognosis across diverse cancer types.^{16 17} In the chemotherapy setting, conventional agents can induce immunogenic cell death (ICD), resulting in re-expressionexpressing or upregulating of tumor-associated antigens (TAAs) by dying malignant cells and subsequently triggering a tumor-reactive CD8⁺ T cell-dependent immune response.^{18 19} The release of TAAs from dying cancer cells during chemotherapy promotes epitope spreading, potentially enhancing both the magnitude and breadth of antitumor immune responses, thereby supporting its use as a complementary partner to ICB. The advent of ICB and high-dimensional profiling approaches, such as single-cell RNA and T cell receptor sequencing (scRNA/TCR-seq), has enabled high-resolution delineation of immune cell subsets within the TME, using robust surface markers to identify the subsets that mediate therapeutic benefits in the context of immunotherapy.20,23 Recent studies have extensively reported that ICB can reinvigorate pre-existing tumor-reactive exhausted T cells or elicit de novo T cell responses, thereby amplifying antitumor response locally and systemically.¹⁷ However, delineating how TME components influence treatment response in the neoadjuvant setting is challenging because chemotherapy is routinely combined with ICB in clinical practice.

We hypothesized that chemotherapy and ICB induce distinct yet overlapping patterns of immune remodeling, and that non-responders retain persistent TAAs with intact antigen presentation machinery, making them ideal targets for TAA-based vaccines. By profiling and comparing paired ESCC tumor samples among treatment arms involving chemotherapy (paclitaxel and cisplatin, TP) or its combination with PD-1 blockade (camrelizumab+albumin-bound paclitaxel and cisplatin/paclitaxel and cisplatin, Cam+nab-TP/TP) using scRNA/TCR-seq, we validated this hypothesis. These findings may facilitate the rational design of novel therapeutics, such as cancer vaccines capable of inducing strong and durable antitumor responses to further improve pCR rate and survival.

Methods

Patients and sample collection

We conducted a multicenter, randomized, open-label phase 3 trial (ChiCTR2000040034) to evaluate the efficacy and safety of neoadjuvant chemoimmunotherapy, compared with chemotherapy alone in patients with ESCC. 55 treatment-naïve patients received two 3-week cycles of neoadjuvant therapy, with 29 patients in Cam+nab TP, 12 in Cam+TP, and 14 in TP group. Surgery was scheduled to occur 4–6 weeks after completion of neoadjuvant therapy. Pretreatment and post-treatment samples were prospectively collected before (n=45) and after neoadjuvant therapy (n=41). Baseline patient characteristics are summarized in online supplemental table S1. In this study, all pathological samples were reviewed by a blind independent review committee, with the tumor regression grade (TRG) evaluated according to Mandard’s TRG scoring system. All samples were collected at Cancer Hospital, Chinese Academy of Medical Sciences, Beijing, China.

Tumor processing for single-cell sequencing

All sample collection procedures followed standard clinical practice. Samples were collected in the RPMI-1640 medium (Gibco), and enzymatically digested with gentleMACS Tumor Dissociation Kit (Miltenyi Biotec) for 60 min on a rotor at 37°C. The dissociated cells were next passed through a 70 µm SmartStrainer in the RPMI-1640 medium (Invitrogen) with 10% fetal bovine serum (FBS) until uniform cell suspensions were obtained. After washing twice with 1x phosphate buffered saline (PBS)(Invitrogen), the cell pellets were resuspended in sorting buffer (PBS supplemented with 1% FBS). scRNA-seq libraries were prepared using the Chromium Next GEM Single Cell 5’ Kit v2 from 10x Genomics, following the manufacturer’s instructions. The libraries were subsequently sequenced on an Illumina NovaSeq 6000 sequencer with 150 bp paired-end reads (Illumina; San Diego, California, USA).

scRNA-Seq data processing

The CellRanger (V.5.0.0, 10x Genomics) was used to aggregate raw data, filter low-quality reads, align reads to human reference genome (refdata-gex-GRCh38-2020-A), assign cell barcodes, and generate the unique molecular identifier (UMI) matrix with default parameters. Quality filtering and downstream analysis were processed with the Scanpy²⁴ pipeline. Specifically, the raw UMI matrix was processed to filter out genes detected in less than 2 cells and cells with fewer than 200 genes or greater than 10% mitochondrial gene counts. We normalized the UMI count matrix to a total of 10 000 counts per cell and then log-transformed the normalized counts (with a pseudo-count of 1). Scrublet²⁵ was used to remove potential doublets in each sample with the expected doublet rate of 0.05.

Dimension reduction and clustering analysis for scRNA-seq data

Highly variable genes were identified based on average expression and dispersion thresholds using the FindVariableGenes function with default parameters, followed by data scaling (ScaleData) and linear dimensionality reduction via principal component analysis (PCA; RunPCA). To correct the batch effects from different patients, the RunHarmony function was run. Clustering was performed on the integrated expression matrix using a shared nearest neighbor graph, and cell clusters were identified at a resolution of 0.4 using FindClusters. Note that the same principal components were also used for non-linear dimension reduction to generate the Uniform Manifold Approximation and Projection (UMAP). After the first round of unsupervised clustering, we annotated each cell cluster according to canonical cell markers. For in-depth investigation, we performed a second round of unsupervised clustering with the same procedure as first round clustering. Cluster-specific marker genes were identified using the Wilcoxon rank-test implemented in the FindAllMarkers function with Bonferroni adjusted p values. The cell clusters were manually annotated according to these marker genes.

TCR sequences assembly

The CellRanger (V.5.0.0, 10x Genomics) was also used to align TCR-seq reads to the human reference genome (refdata-cellranger-vdj-GRCh38-alts-ensembl-5.0.0), followed by TCR sequence assembly, annotation, and basic filtering. Each cell was assigned one alpha beta chain pair with the highest UMI counts. Cells with identical TCR pairs were defined as clonal cells and considered to originate from the same ancestry. The scRepertoire (V.1.8.0)²⁶ package was used for further analysis.

Identification of malignant cells

Copy number variant (CNV) profiles were estimated for each epithelial cell using inferCNV for each therapy group. Briefly, we sorted genes according to their genomic location and calculated average gene expression within each chromosome using a sliding window of 100 genes to deduce the CNV. CD4⁺ T cells from each group were used as a normal karyotype reference for estimation, and CD8⁺ T cells were spiked into the epithelial cells as additional normal cells. Cells were classified as non-malignant only if their assignments clustered together with the majority of CD8⁺ T cells.

Differential gene expression and functional enrichment analysis

Differential gene expressions (DEGs) for each subset versus all other cells were identified using the FindMarkers function in the Seurat R package (V.5.0.0) and filtered using the following criteria: expression in ≥20% of cells in the more abundant group, fold change >1.5, and false discovery rate (FDR) q value <0.05. Ranked gene lists were generated based on log fold-change values, and gene set enrichment analysis (GSEA) was performed using the GSEA function in the clusterProfiler package against the cancer hallmark (H) gene sets. Gene sets with *p<0.05 were considered enriched functional states or procedures.

Cell–cell interaction analysis

To investigate potential interactions between cancer cell and CD8⁺ T cells in the ESCC TME, cell–cell communication analysis was performed by using CellChat²⁷ (V.2.1.2). The default ligand receptor pair information was used in this analysis process, and only receptors and ligands expressed in more than 10% of the cells in the corresponding subclusters were considered.

The ratio of observed to randomly expected cell numbers (Ro/e)

The Ro/e (Ratio of Observed to Expected) is the ratio of observed cell numbers to expected cell numbers within different treatment groups to quantify the enrichment of target cell clusters across the various treatment groups.²⁸ We first applied a χ² test to assess whether these distributions within the treatment groups deviate significantly from random expectations. The expected cell numbers derived from the results of the χ² test. We then calculated the Ro/e for each combination of cell cluster and treatment group using the formula: Ro/e=Observed/Expected. Ro/e serves as an indicator of whether cells from a specific cell cluster are enriched or depleted within a particular treatment group.

Statistical analysis

All statistical analyses and visualization were performed in the R statistical environment (V.4.2.0). The Percent Expressed and Average Expression of TAA and major histocompatibility complex (MHC) genes for each sample was obtained from the data layer of the output of the Dotplot function in Seurat. Statistical comparisons of cell type proportions across different patient groups (eg, Chemo-R vs Chemo-NR) were performed using methods specialized for compositional data analysis. Specifically, cell type percentages were first transformed using the Centered Log-Ratio method via the compositions R package (V.2.0.9) to mitigate the bias inherent in constrained data. Subsequent statistical significance was determined by the Wilcoxon test. For all comparisons, including cell proportions and gene expression levels, p values were adjusted for multiple testing using the Benjamini-Hochberg (BH) FDR method. *Adjusted p<0.05 was considered to indicate statistical significance.

Results

Neoadjuvant therapy induces remodeling of the TME in responders

To decipher dynamic changes in the immune contexture of ESCC in response to antitumor therapies, we prospectively collected paired tumor samples from 55 LA-ESCC patients randomized to receive neoadjuvant chemoimmunotherapy (Cam+nab-TP/TP, n=41) or chemotherapy (TP, n=14) followed by surgery (figure 1A). Using Mandard’s TRG scoring system, patients were divided into responders (TRG1–2) vs non-responders (TRG3–5). The responding rates were 63.41% (26/41) and 35.71% (5/14) for Cam+nab-TP/TP and TP group, respectively (online supplemental table S1). A total of 86 clinical samples were collected pre-neoadjuvant (n=45) and post-neoadjuvant therapy (n=41) and performed scRNA/TCR‐seq. 784 315 high-quality cells were detected following rigorous quality control. Unsupervised clustering based on gene expression profiles identified 10 major cell types for subsequent analysis (figure 1B; online supplemental figure S1A).

Figure 1. Overview of cellular composition of ESCC before and after neoadjuvant therapies. (A) Study design and workflow. Created with BioGDP.com. (B) Uniform Manifold Approximation and Projection (UMAP) visualization of high-quality cells (n=784 315) captured from all enrolled samples (n=86), colored by cell type. Cell types were identified based on the expression of canonical markers and top differentially expressed genes (DEGs) defining each cluster, as shown in online supplemental figure S1A. (C) Proportions of major cell types in pretreatment tumors, stratified by treatment group and clinical response. Combo: Cam+nab-TP/TP; Chemo: TP. (D, E) Comparisons of pretreatment and post-treatment changes in the proportion of major cell types between responders and non-responders from Cam+nab-TP/TP (D) or TP (E) cohort. Boxplots show the first to the third quartile with a line in the middle that represents the median. Whiskers represent 1.5-fold of the IQR. Data were transformed using the centered log-ratio (CLR) method prior to statistical analysis. P values were calculated using the Wilcoxon test and adjusted for multiple comparisons by the Benjamini-Hochberg (BH) method. Cam+nab-TP/TP, camrelizumab plus albumin-bound paclitaxel and cisplatin/paclitaxel and cisplatin; ESCC, esophageal squamous cell carcinoma; Pericyte/SMC, pericyte/smooth muscle cell; scRNA/TCR-seq, single-cell RNA and T cell receptor sequencing; TP, paclitaxel and cisplatin.

Given the well-established association between pretreatment TME characteristics and the effectiveness of anticancer therapies, we first examined whether baseline TME features were correlated with treatment response in our neoadjuvant setting. CD4⁺ and CD8⁺ T cell infiltration was not associated with response to either treatment regimen (figure 1C). Furthermore, Cam+nab-TP/TP responsive tumors harbored a significantly high level of myeloid cell infiltration (*p=0.024). However, this association was not seen in patients treated with TP. No significant differences in the proportions of other cell types were observed between responders and non-responders (figure 1C; online supplemental figure S1B).

Next, we investigated the effect of therapeutic pressure on remodeling of TME in LA-ESCC. Comparative analysis of paired tumors demonstrated that responders from both treatment cohorts exhibited similar dynamic patterns, with post-treatment decreases in CD4⁺ (Cam+nab-TP/TP: ****p=1.2e−07; TP: p=0.064), CD8⁺ T cells (Cam+nab-TP/TP: ****p=9.3e−08; TP: p=0.22) and myeloid cells (Cam+nab-TP/TP: ****p=9.3e−08; TP: p=0.13), along with increases in stromal cell populations, including endothelial Cells, fibroblasts, and pericyte/smooth muscle cells (pericyte/SMCs) (figure 1D,E; online supplemental figure S1C–E). Decreased proportions of epithelial cells were observed after treatment, especially in Cam+nab-TP/TP responders (**p=0.006, figure 1D,E). We then showed that epithelial cells were mainly malignant cells via inferred CNV analysis (online supplemental figure S1H–K). GSEA of malignant cells revealed the downregulation of immune response and proliferation gene programs (online supplemental figure S1F,G), supporting treatment-induced tumor growth control. In contrast, non-responders, irrespective of therapeutic regimen, exhibited a largely static TME with minimal detectable cellular remodeling. Collectively, these findings highlight a clear divergence in treatment-induced TME plasticity, underscoring that only responders undergo meaningful immunologic and stromal reprogramming.

Responders exhibit dynamics in myeloid cell populations

We next examined myeloid cells, an immune compartment that plays a crucial role in antigen presentation and in shaping T cell-mediated antitumor immunity. Analysis of myeloid cells identified 11 subclusters, including five dendritic cell (DC) clusters, two macrophage (Mac) clusters, and four monocyte (Mono) clusters (figure 2A; online supplemental figure S2A). Baseline infiltration of myeloid cell subsets showed no significant differences between responders and non-responders (figure 2B; online supplemental figure S2B). Significant post-treatment changes in myeloid cells were mainly seen in responders. Of particular interest were DCs, all five DC subsets changed in response to Cam+nab-TP/TP. During the therapy, responders exhibited significant increases in cDC1 and cDC2 (**p=0.01 and ****p=7.1e−05), accompanied by marked decreases in pDC-PPP1R14A⁺, pDC-PTCRA⁺, and tDC subsets (**p=0.004, **** p=2.7e−05, and **** p=8.4e−07, figure 2C; online supplemental figure S2C). Although similar trends were observed in TP responders, only a significant increase in cDC2 and a decrease in tDC were detected (*p=0.032 and 0.032, figure 2D; online supplemental figure S2D). In contrast, alterations in myeloid cell proportions were minimal in non-responders and confined to a few subsets, including post-treatment reductions in the pDC-PTCRA⁺ and tDC subsets in the combination cohort (*p=0.019 and p=0.051, figure 2C) and a slight increase in cDC2 in the TP cohort (p=0.051, figure 2D). This indicates that neither neoadjuvant regimen successfully mobilized myeloid cells in non-responders to mount an effective antitumor immune response. Furthermore, neoadjuvant therapies did not compromise interactions between myeloid cell subsets and tumor cells across the cohort (online supplemental figure S2E,F). Overall, these findings highlight dynamic myeloid remodeling as a distinguishing feature of effective therapeutic responses.

Figure 2. Dynamics of myeloid cells in response to neoadjuvant therapies. (A) UMAP clustering and classification of myeloid cells (n=87 652). Subsets were identified based on the expression of canonical markers and top DEG defining each cluster, as shown in online supplemental figure S2A. (B) Proportions of myeloid cell subsets in pre-treatment tumors, stratified by treatment group and clinical response. Combo: Cam+nab-TP/TP; Chemo: TP. (C, D) Comparisons of pretreatment and post-treatment changes in the proportion of myeloid cell subsets between responders and non-responders from Cam+nab-TP/TP (C) or TP (D) cohort. Boxplots show the first to the third quartile with a line in the middle that represents the median. Data were transformed using the CLR method prior to statistical analysis. P values were calculated using the Wilcoxon test and adjusted for multiple comparisons by the BH method. BH, Benjamini-Hochberg; Cam+nab-TP/TP, camrelizumab+albumin-bound paclitaxel and cisplatin/paclitaxel and cisplatin; cDC, conventional dendritic cell; CLR, centered log-ratio; DEG, differentially expressed gene; Mac, macrophage; Mono, monocyte; NR, non-responder; pDC, plasmacytoid dendritic cell; R, responder; tDC, tolerogenic (regulatory) dendritic cell; TP, paclitaxel and cisplatin; UMAP, Uniform Manifold Approximation and Projection.

CD8⁺ T cell dynamics are associated with response to neoadjuvant therapies

We hypothesized that various functional CD8⁺ T cell subsets might be more effectively engaged in responders than in non-responders, since the latter exhibited non-responsive myeloid cells during the therapy. To address this, CD8⁺ T cells were partitioned to 13 subclusters, including CD8⁺ CXCL8⁺, cytotoxic CD8⁺ GZMB⁺, exhausted CD8⁺ GZMB⁺TIGIT⁺ (higher level of TIGIT, PDCD1, and LAG3), effector memory CD8⁺ GZMK⁺, CD8⁺ ISG⁺, CD8⁺ KIR⁺, CD8⁺ MAIT, central memory CD8⁺ Tn/Tmem, CD8⁺ TNFRSF4⁺, CD8⁺ Tprf (proliferation), gd-T, NK and NKT clusters (figure 3A; online supplemental figure S3A). Similar to myeloid cells, baseline infiltration of T cell subsets was not associated with response and showed no significant differences between the two treatment cohorts (online supplemental figure S3B,E). Comparative analysis demonstrated that both Cam+nab-TP/TP and TP responders exhibited post-treatment decreases in the frequencies of GZMB⁺, ISG⁺, and Tprf cells (Cam+nab-TP/TP: **p=0.002, ****p=2e−07, and *p=0.012; TP: *p=0.032, p=0.18 and p=0.064), along with increases in GZMK⁺ and Tn/Tmem cells (Cam+nab-TP/TP: ****p=9.5e−06 and 4.2e−07; TP: *p=0.032 and 0.032, figure 3B,C; online supplemental figure S3C,D). These shifts reflect a transition from chronically stimulated CD8⁺ T cells toward a memory-biased, functionally competent pool and demonstrate that classical chemotherapies can also elicit antitumor immunity, consistent with prior studies.¹⁶ Furthermore, we observed decreased post-treatment GZMB⁺TIGIT⁺ exhausted CD8⁺ T cells in Cam+nab-TP/TP responders. This trend, although not statistically significant, was absent in the TP group. Given the established role of ICB in reinvigorating exhausted T cells,¹⁷ this may represent an ICB-driven effect, with the loss of significance attributed to the limited sample size. To validate the observed phenotypic shift, we next leveraged pseudotime analysis to determine whether the observed changes in CD8⁺ T cell states were supported by an underlying developmental continuum. The trajectory revealed that GZMB⁺ T cells likely originate from an early proliferation state (Tprf), progress through a GZMB⁺TIGIT⁺ exhausted state, and ultimately differentiate into GZMK⁺ effector memory and Tn/Tmem memory states (figure 3D), supporting the aforementioned T cell shift over the course of therapy. GSEA performed on T cell subsets that experienced substantial decreases revealed the downregulation of genes related to immune response after therapy (figure 3E for Cam+nab-TP/TP cohort; online supplemental figure S3F for TP cohort). Thus, neoadjuvant therapy triggers an infection-like CD8⁺ T cell contraction-memory transition, supporting its immunogenic potential. In both scenarios, antigen-experienced effector CD8⁺ T cells undergo a contraction phase during which the majority of effector cells undergo apoptosis, while a small percentage survive to further mature into memory CD8⁺ T cells.²⁹ However, no significant changes in proportions of T cell subsets were observed in non-responders across either treatment cohort (figure 3B,C). These findings indicate that therapy redirects CD8⁺ T cells from early cytotoxic or exhausted phenotypes toward more persistent, long-lived memory populations, suggesting that effective responders undergo immune remodeling that favors durable and robust antitumor CD8⁺ T cell immunity.

Figure 3. Compositional and clonal dynamics of CD8⁺ T cells in response to neoadjuvant therapies. (A) UMAP clustering and classification of CD8⁺ T cells (n=204 941). Subsets were identified based on the expression of canonical markers and top DEGs defining each cluster, as shown in online supplemental figure S3A. (B, C) Comparisons of pre- and post-treatment changes in the proportion of CD8⁺ T subsets between responders and non-responders from Cam+nab-TP/TP (B) or TP (C) cohort. Boxplots show the first to the third quartile with a line in the middle that represents the median. Whiskers represent 1.5-fold of the IQR. Data were transformed using the CLR method prior to statistical analysis. P values were calculated using the Wilcoxon test and adjusted for multiple comparisons by the BH method. (D) UMAP color-coded by pseudotime trajectory analysis of CD8⁺ T cells. (E) Gene set enrichment analysis (GSEA) using the MSigDB Hallmark gene sets on genes ranked by log₂FC (post-treatment vs pre-treatment) in CD8⁺ T cell subsets that significantly decreased after neoadjuvant Cam+nab-TP/TP in pathological responders. Hallmark gene sets were used to compute enrichment scores. Bars show the normalized enrichment score (NES), and the color scale represents the –log10 p value. Statistical significance was assessed using permutation-based p values (10 000 permutations). (F, G) Treatment-induced clonal expansion of CD8⁺ T cell subsets in patients treated with Cam+nab-TP/TP (F) or TP (G). TCR clonotypes were categorized by comparing their abundance between pretreatment and post-treatment samples from the same patient. New: high abundance clones detected only after treatment; Expanded: clones with increased abundance after treatment. Left: responders; right: non-responders. (H, J) Circle plot showing outgoing ligand–receptor interactions from malignant epithelial cells to CD8⁺ T cell subsets in Cam+nab-TP/TP (H) or TP (J) non-responders. Edge width represents interaction strength and node size indicates the number of significant ligand–receptor pairs. (I, K) Bubble plot illustrating predicted ligand–receptor interactions from malignant epithelial cells to specific CD8⁺ T cell subsets in Cam+nab-TP/TP (I) or TP (K) non-responders. The color of each dot represents the interaction probability. Y-axis lists ligand–receptor pairs, and separate panels correspond to different CD8⁺ T cell subtypes. BH, Benjamini–Hochberg; Cam+nab-TP/TP, camrelizumab + albumin-bound paclitaxel and cisplatin/paclitaxel and cisplatin; CLR, centered log-ratio; gd-T, gamma delta T cell; DEGs, differential gene expressions; MAIT, mucosal-associated invariant T cell; NK, natural killer cell; NKT, natural killer T cell; NR, non-responder; R, responder; Tn/mem, naïve/memory T cell; TP, paclitaxel and cisplatin; Tprf, proliferation T cell; UMAP, Uniform Manifold Approximation and Projection.

Next, we investigated the dynamics in cluster distribution of TCR clonotypes in response to neoadjuvant therapy. TCR clonotypes were first categorized by comparing their abundance between pretreatment and post-treatment samples from the same patient. The substantial expansions of pre-existing and new clonotypes were observed in post-treatment tumors (online supplemental figure S3G), again supporting treatment-induced immune responses. Surprisingly, paired TCR clonotype analyses revealed clonality changes in both responders and non-responders, with newly emerged and clonally expanded CD8⁺ T cells mainly originating from GZMK⁺, GZMB,⁺ TIGIT⁺, and Tn/Tmem clusters (figure 3F,G). Cam+nab-TP/TP responders exhibited more pronounced clonal expansion in GZMK⁺ and Tn/Tmem relative to GZMB⁺ TIGIT⁺ T cells compared with non-responders, consistent with the shift observed in their cell frequencies (figure 3F). Moreover, GZMK⁺ and Tn/Tmem clonotypes expanded in parallel with their increased cell proportions (figure 3F,B), indicating the infiltration of diverse newly generated memory T cell clonotypes into the tumor following the combination therapy. These patients therefore, retain the capacity to generate durable immunological memory capable of controlling tumor relapse. Unlike Cam+nab-TP/TP responders, TP responders exhibited the greatest clonal expansion of GZMB⁺TIGIT⁺ cluster but showed minimal changes in its overall abundance (figure 3G,C), mirroring the pattern observed in non-responders. This mismatch suggests that although new exhausted clones emerged, they did not actively engage in the antitumor immunity, resulting in an ineffective T cell response incapable of differentiating to memory T cells.

Given that non-responders exhibited substantial accumulation of tumor-reactive T cell subsets identified in responders, we examined whether their cellular communication with tumor cells was compromised by therapy. Characterizing this interaction is crucial for assessing the potential of therapeutical mobilization of this T cell pool. Cell-cell interaction analysis revealed that stronger communication was mainly seen between malignant epithelial cells and therapy-responsive CD8⁺ T cell clusters, including GZMB⁺, GZMB⁺TIGIT⁺, GZMK⁺, and Tn/Tmem cells in non-responders (figure 3H,J). This result prompted us to examine how neoadjuvant therapy shaped interactions between tumor MHC complexes and predicted TCR. As expected, in non-responders, the majority of recognition pairs persisted on Tprf and GZMB⁺ T cells at the early stage of trajectory, and their interaction strength showed no substantial change (figure 3I,K; online supplemental figure S3H,I). This finding, together with more pronounced clonal expansions of these two subsets in non-responders compared with responders (figure 3F,G), suggests a failure to initiate antitumor immune responses. In contrast, tumors from responders exhibited markedly fewer post-treatment interaction pairs in these two subsets, indicating efficient TCR–HLA recognition and subsequent immune-mediated clearance. New ligand–receptor recognition pairs emerged in memory T cell subsets (GZMK⁺ and Tn/Tmem) within post-treatment tumors, consistent with their increased abundance and clonal expansion. Notably, the CD8⁺ GZMB⁺TIGIT⁺ exhausted cluster displayed the strongest predicted interactions with HLA-A, HLA-B, and HLA-C molecules, as indicated by the red colored high probability signals (figure 3I,K). This suggests that exhausted CD8⁺ T cells maintain robust recognition of classical MHC class I ligands despite their dysfunctional state, consistent with their persistent yet ineffective tumor MHC engagement in non-responders (figure 3F,G).

TAAs persist in non-responders after neoadjuvant therapy

Given that active therapy-induced immune responses were observed in responders and TAAs are well known to elicit T cell-mediated antitumor immunity, we therefore investigated whether and how TAA expression changed in response to neoadjuvant therapy, to better understand their therapeutic relevance. The TAAs examined in the current cohort were selected using public datasets according to strategies summarized in online supplemental figure S4A. After validation in our in-house ESCC cohort using the criteria of higher carrying ratio (>40%), the panel of TAAs was eventually narrowed down to nine candidates, including ACTL8, BRDT, GNGT1, MAGEA1, MAGEA11, MAGEA3, MAGEA4, MAGEA6, and SMC1B. Tumor samples showed striking overexpression of above genes relative to normal tissues (figure 4A). The strong tumor-restricted expression pattern underscores these TAAs as attractive targets for immunotherapy. These genes also exhibited strong and highly restricted expression in ESCC (figure 4B). This tumor-specific expression profile underscores their suitability as clinically actionable targets in ESCC. Responders and non-responders showed no significant differences in baseline TAA expression levels, except for a higher frequency and expression level of BRDT in TP non-responders (online supplemental figure S4B,C).

Figure 4. Expression of tumor-associated antigens (TAAs) in ESCC before and after neoadjuvant therapies. (A) Differential expression of the nine selected TAAs between ESCC tumors and matched normal adjacent tissues. Strategy for the selection is shown in online supplemental figure S4A. *p<0.05, ****p<0.0001 of Wilcoxon rank-sum test. Boxplots display median and IQR. (B) Heatmap illustrating the enrichment of nine selected TAAs in ESCC tumor tissues compared with a spectrum of normal tissues. (C, D) TAA-specific clearance in response to Cam+nab-TP/TP for responders (C) and non-responders (D). Heatmaps visualize the percentage of TAA-expressing cells (left) and the average TAA expression level (right). Top: pretreatment; bottom: post-treatment. (E, F) TAA-specific clearance in response to TP for responders (E) and non-responders (F). Cam+nab-TP/TP, camrelizumab+albumin-bound paclitaxel and cisplatin/paclitaxel and cisplatin; ESCC, esophageal squamous cell carcinoma; NR, non-responder; R, responder; TP, paclitaxel and cisplatin; TPM, transcripts per million.

Next, we assessed changes in TAA expression following treatment. In Cam+nab-TP/TP responders, only 15.8% (3/19) post-treatment tumors remained TAA-positive, whereas TAAs were still detectable in the majority (66.7% (8/12)) of non-responders (figure 4C,D). Responders showed significant reductions across all nine TAAs (online supplemental figure S4B). These findings indicate effective treatment-driven elimination of TAA-expressing tumor cells in responders, in sharp contrast to the persistence of these antigens in non-responders. Similarly, TP responders also showed a clear TAA-specific clearance after treatment, with post-treatment tumors exhibiting nearly complete loss of detectable TAA expression (figure 4E,F; online supplemental figure S4C). Collectively, treatment-induced clearance of TAA-high tumor clones in responders supports the high immunogenicity of these TAAs, meaning that they should be easily recognized and targeted by immune responses. Thus, their high prevalence and abundant presence in non-responders make them top candidates for novel therapeutics.

HLA expression is not compromised over the course of neoadjuvant therapy

The persistence of TAAs in non-responders prompted an investigation into whether neoadjuvant therapy impaired antigen processing and presentation machinery (APM) in LA-ESCC patients. A panel of HLA class I genes including the classical antigen-presenting genes HLA-A, HLA-B and HLA-C, as well as the non-classical genes HLA-E, HLA-F, and HLA-G, was analyzed, along with key components of APM (B2M, CALR, PDIA3, TAPBP, TAP1, and TAP2). Responders and non-responders showed highly similar HLA-I/APM expression patterns before and after treatment, indicating that APM defects do not explain primary resistance in either treatment cohort (figure 5A, B, E and F). Figure 5C,D,G,H further confirms that APM remained largely intact in the current cohort. Specifically, neither group exhibited substantial reductions in percent-expressing cells or average expression levels of HLA class I molecules and key APM components following therapy. Despite observing decreased expression in a limited number of genes (CALR, TAP1, and TAP2: p=0.072, 0.072, and 0.11, figure 5D) in Cam+nab-TP/TP responders, these shifts remained modest and did not indicate pathway impairment. Importantly, the absence of any HLA-I/APM shift in non-responders indicates fully preserved APM function despite treatment failure, suggesting that treatment resistance is not driven by loss of antigen presentation, but rather by the failure to eliminate TAA-positive tumor clones.

Figure 5. Expression of MHC class I genes in ESCC before and after neoadjuvant therapies. (A, B) Heatmap showing expression of MHC class I genes in pretreatment and post-treatment tumors from Cam+nab-TP/TP responders (A) and non-responders (B). Top: pretreatment; bottom: post-treatment. Left: percentages of tumor cells expressing each gene; right: average expression levels of each gene. (C) Comparisons of pretreatment and post-treatment changes in the percentages of tumor cells expressing each MHC class I gene between responders and non-responders in the Cam+nab-TP/TP cohort. Boxplots show the first to the third quartile with a line in the middle that represents the median. Whiskers represent 1.5-fold of the IQR. P values were calculated using the Wilcoxon test and adjusted for multiple comparisons by the BH method. (D) Comparisons of pre- and post-treatment changes in the average expression level of each MHC class I gene between responders and non-responders in the Cam+nab-TP/TP cohort. (E, F) Heatmap showing expression of MHC class I genes in pre- and post-treatment tumors from TP responders (E) and non-responders (F). Top: pretreatment; bottom: post-treatment. Left: percentages of tumor cells expressing each gene; right: average expression levels of each gene. (G) Comparisons of pretreatment and post-treatment changes in the percentages of tumor cells expressing each HLA class I gene between responders and non-responders in the TP cohort. (H) Comparisons of pretreatment and post-treatment changes in the average expression level of each MHC class I gene between responders and non-responders in the TP cohort. BH, Benjamini-Hochberg; major histocompatibility complex; Cam+nab-TP/TP, camrelizumab+albumin-bound paclitaxel and cisplatin/paclitaxel and cisplatin; ESCC, esophageal squamous cell carcinoma; NR, non-responder; R, responder; TP, paclitaxel and cisplatin.

Similar results were obtained when we examined HLA class II expression in both Cam+nab-TP/TP and TP cohorts (online supplemental figure S5). Of note, we observed an increase in the percentage of tumor cells expressing CD74 after Cam+nab-TP/TP treatment (p=0.11, online supplemental figure S5C). The same trend was also noted when we examined TP responders. The neoadjuvant treatment induced a factor that may affect prognosis,³⁰ while it may also serve as a potential therapeutic target for which drugs are available³¹ in the postoperative management. Overall, neoadjuvant therapy does not compromise expression levels of the majority HLA genes, consistent with the observed T cell clonal expansions and previous findings.³²

In summary, our data demonstrate that chemotherapy and immunotherapy work synergistically to remodel the TME and ultimately clear TAA-high tumor clones in responders, characterized by restructured DC and CD8⁺ T cell subsets and a pronounced shift from epithelial to stromal compartment. In contrast, TME characteristics in non-responders include TAA persistence, deactivated DCs, dysfunctional GZMB⁺TIGIT⁺ T cell expansion, and intact HLA machinery, pointing to the potential of a combined vaccination strategy integrating TAA prediction, DC loading, and ICB (figure 6). We therefore compile these findings into a comprehensive translational framework (figure 7), offering an integrated overview of the above TME components in pretreatment and post-treatment tumors across treatment groups. This informative framework could facilitate a rapid understanding of patient TME properties and is helpful for appropriate decision-making in clinical practice of tumor control.

Figure 6. Dynamics of TME in responders and non-responders following neoadjuvant therapies. Responders exhibit active antitumor responses characterized by dendritic cell (DC) activation, exhausted CD8⁺ T cell contraction, and memory T cell expansion, leading to cancer elimination. Conversely, non-responders show a GZMB⁺TIGIT⁺ T cell expansion and persistent expression of targetable TAAs with intact HLA machinery. These vulnerabilities could be targeted by a comprehensive triplet combination vaccination strategy that integrates TAA/neoantigen prediction, DC loading, and PD-1 blockade. Created with BioGDP.com. HLA, human leukocyte antigen; PD-1, programmed cell death protein 1; PD-L1, programmed cell death-ligand 1; TAA, tumor-associated antigen; TCR, T cell receptor; TME, tumor microenvironment.

Figure 7. Summary of individual TME characteristics before and after neoadjuvant therapies in the entire ESCC cohort. Cam+nab-TP/TP, camrelizumab + albumin-bound paclitaxel and cisplatin/paclitaxel and cisplatin; cDC, conventional dendritic cell; ESCC, esophageal squamous cell carcinoma; Mac, macrophage; MAIT, mucosal-associated invariant T cell; Mono, monocyte; NR, non-responder; pDC, plasmacytoid DC; Pericyte/SMC, pericyte/smooth muscle cell; R, responder; TAA, tumor-associated antigen; TME, tumor microenvironment; MHC, major histocompatibility complex; Tn/mem, naïve/memory T cell; TP, paclitaxel and cisplatin; Tprf, proliferation T cell; tDC, tolerogenic (regulatory) DC.

Discussion

While neoadjuvant ICB has been widely used in the real world patient population, the pCR rate remains low. Novel therapeutic agents are urgently needed to improve response rates, and their development largely relies on identifying immunological correlates of immune response. We assessed intratumoral immune changes following treatment using longitudinal sampling from LA-ESCC patients treated with neoadjuvant chemotherapy with or without PD-1 blockade and demonstrated the following: (1) responders exhibit an active antitumor response closely resembling that elicited during an infection, regardless of regimen; (2) multiple immune cell subsets change significantly following neoadjuvant therapy, whereas the suppression of GZMB⁺TIGIT⁺ CD8⁺ T cell clonal expansion was observed only in responders to chemoimmunotherapy; and (3) a panel of TAAs validated in post-treatment non-responders shows the potential to induce antigen-specific antitumor immunity.

Our data revealed similar temporal dynamics of immune characteristics associated with clinical response in Cam+nab-TP/TP and TP cohorts, though cohort-specific variations were observed. The similarity is not surprising because potent chemotherapy agents have been reported to increase the immunogenicity of malignant cells by potentiating both antigenicity (via TAAs) and adjuvanticity.¹⁶ Such “viral mimicry” mode of killing neoplastic cells is also a mechanism underpinning the efficacy of other modalities including targeted therapy and radiotherapy.³³ In a recent scRNA-seq study of patients with rectal cancer treated with total neoadjuvant therapy (TNT, consisting of short-course radiotherapy (nRT, 5×5 Gy) followed by 6–8 cycles of FOLFOXIRI chemotherapy), the observed increase in CR is partly attributed to the synergistic cytotoxic effects resulting from radiotherapy-induced DNA damage and the in-field and spatial cooperative effects of combining it with chemotherapy.³⁴ Consistent with our findings, TNT efficacy is associated with increased IFNG⁺CD8⁺ effector memory T cells with high expression of activation marker genes including GZMK. Importantly, the proportion of this particular T subset decreased after nRT but increased after TNT in 4 patients with matched Pre, nRT, and TNT samples. These TME remodeling patterns are consistent with the established cell death signatures mediated by different treatments. Radiotherapy has been reported to induce ICD and the release of damage-associated molecular patterns (DAMPs).³⁵ However, this process is often accompanied by immunosuppressive effects resulting from radiation-induced lymphodepletion.^{34 36 37} Thus, this dual effect may contribute to the suboptimal short-term efficacy of concurrent chemoradiotherapy combined with immunotherapy in ESCC.³⁸ On the contrary, chemotherapy can induce a complete ICD signature encompassing both antigenicity and adjuvanticity, which can be assessed by its core hallmarks including expression levels of DAMPs (eg, ATP, HMGB1, and ANXA1), phosphorylated eIF2α or related stress response gene expression, cell surface exposure of “eat-me” Signal CALR, activation of danger signaling mediated by type I IFN and T cell recruitment chemokines such as CXCL10.³⁹ These signals collectively activate DCs, promote cross-presentation of tumor antigens, and initiate antitumor immune responses. As demonstrated in the present study, TP induced DC remodeling and clonal expansion of CD8⁺ T cells, including GZMB⁺ TIGIT⁺ subset representing exhausted tumor-reactive T cells that have been established as ICB-responsive in multiple scRNA-seq studies.40,42 This result suggests chemotherapy-induced ICD,^{33 43 44} enabling subsequent local recognition of tumor antigens. However, although not statistically significant, a post-treatment reduction in the proportion of GZMB⁺TIGIT⁺ T cells and the suppression of their clonal expansion were observed exclusively in Cam+nab-TP/TP responsive tumors. This finding suggests the established role of ICB in reinvigorating exhausted T cells and is consistent with other ESCC scRNA-seq studies.^{45 46} This underscores another critical determinant of productive ICD, that is, permissive microenvironmental conditions that facilitate the recruitment and functional activity of T cells. In the present cohort, GZMB⁺TIGIT⁺ CD8⁺ T cells were more abundant than GZMB⁺ cells, suggesting that the immunosuppressive tumor microenvironment strongly impedes chemotherapy’s efficacy as an ICD inducer. Our data and previous studies indicate that while chemotherapy alone demonstrates very limited therapeutic efficacy, it is particularly effective at driving adaptive immunity, making it a highly suitable combination modality. Lack of response has been ascribed to limited antigenicity on tumor cells⁴⁷ and immune evasion mediated by dysfunctional state of antitumor T cells consequent to chronic antigen stimulation during cancer progression.⁴⁸ Within this concept framework, we propose that chemotherapy-induced ICD and ICB-induced cytotoxic revival of exhausted T cells work synergistically to enhance immune response, potentially explaining the higher pCR rate provided by the chemoimmunotherapy compared with chemotherapy, as has been previously reported.^{18 49}

In comparison to previous scRNA-seq studies, we leveraged the advantage offered by this high-resolution technique to examine how TAA expression on tumor cells changed in response to the treatment, which cannot be achieved using traditional bulk RNA-seq. Therapy-induced efficient clearance of TAA-positive tumor cells in responders suggests their antigenicity to elicit antitumor response, keeping in line with previous studies.^{41 50} Notably, we confirmed stable TAA identity and expression level during neoadjuvant therapy across non-responders. These results, along with uncompromised HLA expression on therapy, suggest that TAAs may represent feasible targets to further enhance tumor-specific T cell response.

Our biological findings highlight the significance of potentiating the antitumor immune response to enhance therapeutic efficacy. Immune characteristics in both responders and non-responders provide critical insights for this purpose, specifically for targeting critical immune cell subsets and augmenting tumor antigen exposure. Responders exhibited dynamics of multiple DC clusters, which did not occur in non-responders. Consistently, both pre-clinical and clinical studies have shown that DCs are key mediators of cellular cross talk in the TME^{51 52} and are positively correlated with favorable prognosis.53,55 In-depth scRNA-seq analysis of both tumor tissues and peripheral blood in patients with squamous cell head and neck cancers showed decreases in multiple DC subsets during ICB,⁴¹ suggesting their importance in both local and systemic tumor immunity. Similarly, analysis of the single-cell transcriptional landscape of esophageal adenocarcinoma showed that the ratio of conventional to plasmacytoid DCs markedly reduced in tumor but increased following neoadjuvant chemotherapy,⁵⁶ consistent with our findings. Therefore, treatment-responsive cDC2 and pDC may be essential targets in efforts to generate therapeutic immunity against cancer in the context of ESCC. The strategy combining the chemoattractive properties of pDCs with the superior T cell priming properties of cDC2s in a single cancer vaccine⁵⁷ has already been tested in prostate cancer patients (NCT02692976)⁵⁸ and metastatic endometrial cancer (NCT04212377). Furthermore, we identified GZMB⁺TIGIT⁺ CD8⁺ T cells as treatment-responsive T cells in Cam+nab-TP/TP responders, suggesting their tumor-reactive nature and also highlighting the necessity of incorporating ICB in combination strategies. However, this cell subset in Cam+nab-TP/TP non-responders showed no response to ICB. The residual cytotoxic functionality of these cells cannot be unleashed through the removal of inhibitory signals, suggesting that ICB resistance is mainly driven by the functional exhaustion of T cells rather than their absence. Further research is needed to elucidate the underlying mechanisms. Importantly, non-responsive tumors also provide a valuable opportunity to examine TAA status following antitumor therapies. We set up a cut-off of prevalence rate at 40% to select candidate TAAs, which is crucial for the development of shared cancer therapies aiming at targeting sufficient proportion of patient population, enabling substantial resource and time savings in therapeutic design. Moreover, we did not observe significant downregulation of HLA class I/II genes in non-responders, suggesting HLA loss-of-function mutations or loss of heterozygosity are unlikely to be the primary resistance mechanism in this cohort. Collectively, we propose a comprehensive triplet combination strategy that combines TAA/neoantigen prediction, DC loading, and ICB. Future studies will prioritize MAGE-A3/A4-derived HLA-A*02:01-restricted epitopes with high predicted binding affinity (NetMHCpan EL%rank<0.5) for vaccine design.

Our study is limited by the sample size of each cohort and imbalance between cohorts. TP arm consists of only 14 patients (five responders) compared with 41 (26 responders) in Cam+nab-TP/TP arm. Nevertheless, clear trends with marginal statistical significance were observed even within our limited sample set. Multicenter validation cohorts are underway to confirm these immune signatures as predictive biomarkers for neoadjuvant response. The observed clearance of TAA-positive tumor cells following therapy provides indirect evidence for a TAA-induced T cell response. Although we did not perform MHC tetramer staining or in vitro killing assays, the clonal expansion of CD8⁺ T cells targeting TAA-expressing tumor cells strongly suggests antigen-specific cytotoxicity. Our preclinical functional validation and clinical studies focusing on their therapeutic potential to enhance immune response in ESCC are ongoing, which will provide valuable opportunity to validate our findings and extend the understanding of therapeutic-induced tumor immunity.

In conclusion, we comprehensively investigate how TME is remodeled by neoadjuvant chemotherapy with or without PD-1 blockade, uncovering both shared and ICB-distinct TME characteristics associated with treatment response. Moreover, our data demonstrate that non-responders exhibit clonal expansion of exhausted GZMB⁺TIGIT⁺ CD8⁺ T cells and persistent TAA expression with intact HLA machinery, representing an optimal cohort for TAA-targeted therapeutic vaccines in combination with ICB. The study lays the foundation for leveraging single-cell data to guide the development of novel therapeutic vaccines to optimally mobilize antitumor immunity.

Data availability statement

Data are available on reasonable request.